Word error rate calculator

Word error rate is the standard measure of transcription accuracy and one of the most widely misquoted numbers in speech technology. Two properties account for most of the trouble: it is not a percentage in the ordinary sense, and it is meaningless without the normalisation that produced it.

The formula, and why the denominator matters

WER is the number of substitutions, deletions and insertions required to turn the hypothesis into the reference, divided by the number of reference words. Those three counts are not observable directly — they come from aligning the two word sequences by minimum edit distance, which is why any honest WER tool has to compute an alignment before it can report a number, and why this one shows you that alignment.

Because insertions are unbounded but the denominator is fixed, WER can exceed 100%. A recogniser that hallucinates a paragraph onto a short utterance can score 500%, and this is not a defect in the measure so much as a consequence of what it counts. The practical implication catches people out constantly: word accuracy is not 1 − WER in general. Past 100% that subtraction produces a negative accuracy, so this calculator declines to report one and shows the match error rate instead.

WER vs CER

Character error rate runs the same alignment over characters rather than words, and the difference is partial credit. "Sleeping" transcribed as "sleepin" is a whole substitution at word level and a single character deletion at character level, so CER is usually the lower of the two. That makes CER more informative when you care whether output is readable rather than exact, when the language is morphologically rich, and when word boundaries are not marked in the writing system. Neither is more correct; they answer different questions, and both are reported here from the same normalised text so they describe the same comparison.

The two other measures on this page come from Morris, Maier and Green (2004), who proposed the match error rate and word information lost precisely because WER's unbounded range makes system comparison awkward. Both are bounded by 100% by construction.

The normalisation is part of the result

Case, punctuation, filler words and speaker labels each move WER by several points. A transcript that writes "it's" where the reference has "its" is wrong under a strict comparison and identical under a lenient one, and neither convention is more legitimate — they answer different questions about whether that distinction matters downstream. What is not defensible is reporting the number without the convention, because two WERs computed differently are not comparable even for the same audio and the same system. The toggles on this page are deliberately in front of the result rather than behind a settings panel, and the copy button writes the list of applied normalisations into the output.

What a single number hides

Two systems at 12% WER can be failing in ways that need entirely different remedies. One may be dropping quiet or overlapping speech — a microphone and diarisation problem. Another may be substituting acoustically similar words, which is a language-model or domain-vocabulary problem. A third may be inventing fluent text with no acoustic basis, which is the failure mode neural transcription introduced and the one most dangerous in a clinical record, because the output reads perfectly.

The diff distinguishes them at a glance, and this tool flags the two lopsided patterns explicitly: insertions heavily outnumbering deletions usually means hallucinated or duplicated text, and the reverse usually means truncated audio or a skipped segment. Our guide to preparing audio and speech for annotation covers the upstream decisions — diarisation, overlapping speech, verbatim versus clean conventions — that determine which of these you will be looking at.

Where transcription error goes next

For most research uses the transcript is an intermediate artifact, and the question is not its WER but whether the errors it contains change the coding done on it. A substituted filler word rarely does; a dropped negation or a misheard symptom term can flip an item rating. That makes transcription error one input to a label-quality audit rather than a separate concern — our note on finding label errors in a dataset treats the same problem from the annotation side, and if you are budgeting the correction pass, the transcription time calculator estimates the hours that AI-plus-review actually takes at a given accuracy.

Once the transcript is stable, the reliability questions move to the coding on top of it: whether two coders agree, which the inter-rater reliability calculator answers for categorical codes, and whether the instrument's items hang together, which is what the Cronbach's alpha calculator is for. Interview-based instruments such as the MADRS are scored from exactly this kind of transcript, which is why an error rate on it is worth knowing before the ratings are trusted.

What this tool does not do

It does not normalise numbers, dates or currency, so "twenty five" against "25" scores as a substitution. Doing that properly is language-specific and full of judgement calls, and a half-implemented version would silently change scores in ways users could not predict. It does not compute a per-speaker breakdown, and it does not diarise: if your transcripts label speakers, strip the labels or leave them in on both sides, but the tool will not attribute errors to one speaker.

It also aligns exactly rather than approximately, which is the right trade for accuracy and the wrong one for very large inputs. Two book-length transcripts with many differences will exceed the memory budget and the tool will say so rather than freezing the tab. Compare a section at a time; a WER computed on a representative sample is more useful than one that never finishes.

Frequently asked questions